Identifying Similar Eye Movement Patterns with t-SNE

Abstract

In this paper we describe an approach based on the t-distributed stochastic neighbor embedding (t-SNE) focusing on projecting high-dimensional eye movement data to two dimensions. The lower-dimensional data is then represented as scatterplots reflecting the local structure of the high-dimensional eye movement data and hence, providing a strategy to identify similar eye movement patterns. The scatterplots can be used as means to interact with and to further annotate and analyze the data for additional properties focusing on space, time, or participants. Since t-SNE oftentimes produces groups of data points mapped to and overplotted in small scatterplot regions, we additionally support the modification of data point groups by a force-directed placement as a post processing in addition to t-SNE that can be run after the initial t-SNE algorithm is stopped. This spatial modification can be applied to each identified data point group independently which is difficult to integrate into a standard t-SNE approach. We illustrate the usefulness of our technique by applying it to formerly conducted eye tracking studies investigating the readability of public transport maps and map annotations.